Getting LLM to think and act like a human being: logical path reasoning and replanning

Large Language Models (LLMs) have significant reasoning capabilities and can act as agents interacting with the real world. However, they are often segmented and, unlike humans, lack integrated systems for validating their thoughts and actions. This limitation often leads LLMs to encounter “local optima” in task performance. To mitigate this problem, we propose a replanning mechanism for LLM-based agents that dynamically incorporates feedback from actions and exploits implicit information not initially available in the reasoning framework. This approach effectively bridges the gap between the cognitive and action phases of LLMs. Experimental results on real-world ticket booking platforms such as Ctrip.com and Booking.com show that our method exhibits greater robustness in following clear instructions, successfully completing more steps, and achieving a higher success rate in practical applications, especially in complex tasks requiring interactive reasoning and action.